BEN EBSWORTH
INFRA | SOFTWARE | HARDWARE
Latest writing
All posts →Engineering· 10 min read
Beating Splash on an M5 Max: where the tokens per second actually were
A Rust and candle inference engine went from 0.56x to 1.2x Splash's decode speed on a 27B model. The kernels were never the problem. The GPU sat waiting for the CPU.
Engineering· 8 min read
Astra extra high vs SWE-2: lighthouses and racing cars
GPT-6 Astra at extra-high reasoning and Devin SWE-2 High run the same eleven tasks, including a rotating voxel lighthouse and a car racing game.
Engineering· 9 min read
The behavioural scorer caught a model lying about its own game
We wired three frontier-class models (Gemini 3.6 Flash, Claude Opus 4.6 thinking, and GPT-OSS 120B) into the same 7-task harness via the Agy CLI, then switched the scorer from HTML structure to Playwright behavioural checks. The headline result: a model that scored a perfect 100 on the platformer task under the old scorer scored 30 under the new one, five iterations in a row, because the Space key never actually jumped.
From the lab
All 30+ effects →Small, working simulations. Drag the controls on the full pages, or just watch the previews cycle here. The selection rotates daily.